Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
AI based Microscopy diagnosis of MalariaAI for Health Workshop 2019
Rose Nakasi
Makerere UniversityArtificial intelligence and data science Lab
2/9/2019
1 / 24
Microscopy Diagnosis
Figure: Gold standard for diagnosis of malaria is a microscope2 / 24
Health diagnostic burden
Currently in Uganda, and many developing countries
• Malaria has beenreported as one of theleading cause of deathaccounting for over27% of lives ofUgandans
• Patient in big numberswait to be diagnosed.
3 / 24
Currently in Uganda, and many developing countries
• Most of these countriesdon’t have enoughtrained lab technicians.
• In Ghana, 1.72microscopes per100,000 population,but only 0.85 trainedlaboratory staff per100,000 population .
4 / 24
Microscopy Diagnostic challenge
Currently in Uganda, and many developing countries
• SOP requires not toview more than 20slides a day
• IMicroscopy is eyestraining
5 / 24
Concept detail
Figure: Adapter set up and attachment.
6 / 24
Concept detail
7 / 24
Concept detail
8 / 24
Captured image
Figure: Malaria microscopic image captured with a smartphone.
9 / 24
Plasmodium detection accuracies
Figure: Plasmodium detection results.
10 / 24
ROC AUC for malaria, TB and intestinal parasitesdetection
Figure: Deep learning ROC accuracies.
11 / 24
Detected Objects for plasmodium (left) andbacilli(right).
12 / 24
Server end app.
13 / 24
Server end app.
Figure: Before detection.
14 / 24
Server end app.
Figure: After detection.15 / 24
Framework assessment requirements
Reasons to consider for an effective clinical framework;
• Representativeness in data
• Explainability
• Potential biases in the data.
16 / 24
Representativeness in data
Training data is from a different context;
• Obvious case: do not use European training data for medicaldiagnostics in East Africa.
• Less obvious case: Medical diagnostics data could usetraining data from East Africa for an algorithm in WestAfrica – may under-diagnose malaria.
Training data is imbalanced;
• Data that is not balanced in terms of pixel dimensions,shape, color etc may bias certain algorithms to favor certaingroups
17 / 24
Explainability
Ethical coonsiderations;
• Do Organizations think of ethics when using data/outcomesfrom algorithms, resulting organizational values not beingcaptured in outcomes.
Relevancy / Value;
• Should organizations be using AI and does AI provide enoughvalue that offsets potential complications?
Transparency;
• How transparent are the algorithms/outputs?
18 / 24
Potential data Biases
Does your data have biases reflected in it? If so, your algorithms mayamplify these biases;
• Due to poor annotations.
• Different image dimensions.
• Different environments under which the images werecaptured.
19 / 24
What can we do?
• Auditable algorithms
• Implement fairness
• Accountability and responsibility
20 / 24
Auditable algorithms
Bias is introduced at the algorithm level
• Developers are making decisions to choose between equity,accuracy, speed
Ensure that algorithms are auditable;
• Can they be monitored by external actors? How can you testthe algorithms?
21 / 24
Fairness
Data will never be perfect. How will we deal with imperfect data?
• Getting better data
• Implementing fairness/equity at the algorithm level
22 / 24
Accountability and responsibility
What mechanisms can we add to make sure that we are holding ourselvesand algorithms accountable?
• We work in complex situations and downstream effects arehard to predict
• Build in monitoring to make sure that we can adapt to andfix inequalities that result
• Above all, an AI Assessment Framework to guide AI forhealth Solutions.
23 / 24
Thank you!
Questions?
24 / 24